1-Implement the perceptron learning algorithm for 2 inputs and show graphically how the separating line adjusts its position in the plane during weight updates (animation!).
[url removed, login to view] how the perceptron successfully learns to distinguish linearly separable data sets.
[url removed, login to view] how the perceptrons fails to successfully learn to distinguish linearly non-separable data sets.
2-Implement the basic error backpropagation algorithm for multilayer perceptrons. For testing purposes, choose as a training and test set a unit circle and a two-class problem ‘inside the circle’ and ‘outside the circle’ and give a graphical representation of the decision surface, the training data as well as test points. Be very careful when implementing the algorithm; code for neural networks is notoriously difficult to debug!
3-There exists a database on the Internet with real world data for machine learning; you can access the database at [url removed, login to view] Choose one data set you like and run your neural network program on it. Report statistically significant results.
4-Implement weight decay as a technique for improving the generalization performance of trained networks.
5-Using your Implementation of the basic error backpropagation algorithm for multilayer perceptrons (or a downloaded version of the program if your own code did not work), to show how prior knowledge can be used to initialize a neural network prior to training and what the effects that prior knowledge has on training and generalization performances. Use a toy problem to demonstrate that your network trains faster with prior knowledge compared to training without prior initialization.
6-You will apply the methodology of knowledge-based neural networks discussed in class to a real-world problem from molecular biology. The text that gives further details on the methodology and the application is attached. There exists a database on the Internet with real world data for machine learning; you can access the database for this project at ftp://[url removed, login to view] Run your neural network program on it and report statistically significant results.